The future of SEO, a practical guide to RAG, AI search and the new rules

Home / Blog / The future of SEO

Published 9 October 2026. Every factual claim here was verified against primary documentation on that date, and sources are named in the text.

This article avoids prediction. The topic is full of confident forecasts and invented statistics, so everything below is either documented by Google, documented by a named source, or clearly labelled as our own model. Where the honest answer is that nobody knows yet, it says so.

What has actually changed, and what has not

The most useful thing you can do with this topic is separate the two columns below, because most of the anxiety in this market comes from treating everything as new.

Google states the position directly in its AI optimisation guidance: its AI features rely on core Search ranking and quality systems, so standard SEO practices still apply, and it treats the newer labels AEO and GEO as SEO aimed at the AI search experience rather than as separate disciplines. That single sentence should settle a lot of internal debates.

How RAG works, and why it matters to your pages

RAG stands for retrieval-augmented generation, and understanding it in one paragraph changes how you write. A language model on its own produces text from patterns it learned during training, which is why it can sound confident and be wrong. RAG adds a step before generation: the system retrieves relevant current material, then generates its answer grounded in that material, usually citing some of it.

The practical consequence is that there are now two separate competitions, and most businesses only think about the first.

1. The retrieval competition. Does your content get fetched as source material for the answer. This is where crawlability, indexing, relevance and entity clarity decide your fate.
2. The citation competition. Once retrieved, does your page get named or linked in the generated answer, or does it silently inform an answer that credits someone else.

You can win the first and lose the second, which is the most frustrating outcome and the hardest to see in reporting. It is also why the old instinct of writing one long page covering everything has weakened. If a system retrieves passages to answer a specific sub-question, a page with a clear, self-contained answer to that exact sub-question is easier to use than a page where the answer is spread across six paragraphs.

Query fan-out, the single biggest shift

This is the most important mechanical change and the least understood. Google's documentation states that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources, and that while a response is generated the systems identify additional supporting pages, which lets Google show a wider and more diverse set of links.

Read that again, because it inverts a core assumption. You are no longer optimising for the query a person typed. You are competing for the dozens of sub-queries the system generates on their behalf, most of which you will never see in any keyword tool.

A worked illustration. Someone asks an assistant to help them choose a supplier for an office fit-out in Dubai. Behind that one question, a fan-out might run searches on typical fit-out stages, approvals needed in a Dubai commercial tower, how fit-out contractors are compared, typical timelines, what to ask before signing, and which firms operate in that area. Your beautifully optimised "office fit-out Dubai" page competes for one of those. The firm that has also published a clear answer on approvals, on timelines and on what to ask before signing has six entry points instead of one.

This is why topic coverage now beats keyword targeting, and why the practical instruction is unglamorous: write the obvious adjacent questions your sales team answers every week, each as its own clearly headed answer. Our guide on traffic that does not become enquiries covers how to tell whether that coverage is reaching commercial pages or just producing readers.

The RAISE framework, our model for AI-era SEO

Checklists for AI search are either vague or invented. So we organised Google's published guidance into five layers that can be audited in order, because each one is useless without the one above it. This is Tomsher's own model for structuring the work. It is not an industry standard and Google does not endorse it; every requirement inside it traces back to documented guidance.

Audit in order. Most businesses we review fail at R or I, then spend their budget on S, which is why their content programme feels like it is not working.

New terms and concepts explained

The vocabulary around this topic is being used loosely, often by people selling something. Here is the working set with what each term actually means.

Five things Google says you can skip

This is the most commercially useful section in the article, because each item is currently being sold somewhere in this market. All five come from Google's own AI optimisation guidance.

1. llms.txt and other special AI markup. Google says Search does not use them. The clearest on-record statement came from Google's John Mueller on 17 June 2025: "FWIW no AI system currently uses llms.txt." Google's AI optimisation guidance now lists it among the things you can skip. It costs almost nothing to publish, so treat it as an unproven experiment rather than a deliverable worth paying for.
2. Chunking your content into small pieces. Google says there is no ideal page length and this is not necessary.
3. Rewriting your content specifically for AI systems. Listed explicitly as something you can skip.
4. Chasing inauthentic mentions across the web. Google says this is not as helpful as it seems, which is worth remembering next time someone offers a package of brand mentions.
5. Over-focusing on structured data. Not required for AI search, though it still helps with rich results, so implement it for the right reason rather than as an AI tactic.

One more line from the same guidance is worth putting in front of anyone buying an AI visibility tool: "No third-party tool has access to our internal ranking or AI systems." Monitoring tools are useful for observing how assistants describe your brand. None of them can report a score from inside Google.

How to measure AI search visibility

Until recently this was guesswork. It is not any more, and most businesses have not noticed the change.

Search Console now includes a Generative AI performance report, which Google says rolled out to all websites worldwide as of 31 August 2026. It shows impressions from AI Overviews and AI Mode, where an impression means how many times links to your site were shown to a user in a generative AI feature on Google Search. You can group the data by page, country, date or device, and filter by text-based or multimodal searches such as Google Lens.

Four caveats that matter before you build a report on it. Data only appears once your site has enough impressions in those features. Search Labs experiment data is excluded. The newest data is preliminary and can change within hours. And it inherits the Search performance report's limits, including the 1,000-row cap, with dates in Pacific Time.

Used sensibly, this gives you something the whole industry lacked a year ago: a first-party number for AI visibility, from the platform itself, rather than a third-party estimate. Pair it with the reporting discipline in our SEO audit work, and with the tool comparison in our guide to 30 SEO tools for the monitoring that sits outside Google.

The controls you now have

Businesses regularly ask whether they can decide how their content is used. You have more control than you probably realise, and the controls do different things.

A practical note from Google's documentation: if content still appears in AI features after you add a preview control, confirm the control is correct and visible to Googlebot using the URL Inspection tool, and allow time for recrawling, which can take days to months.

Our own view, for what it is worth. Excluding yourself from AI features to protect clicks is a decision we would challenge hard before implementing, because you lose the impressions and the brand presence while competitors keep theirs. There are legitimate reasons to do it, and preserving traffic is rarely one of them.

Agentic commerce and the machine-readable web

This is the part that will matter most in two years and is almost absent from the SEO conversation today. The shift is from systems that read your site to systems that transact with it.

On 11 January 2026, Google announced the Universal Commerce Protocol, an open-source standard for agentic commerce developed with Shopify, Etsy, Wayfair, Target and Walmart, and endorsed by more than twenty partners including Adyen, Mastercard, Stripe and Visa. The problem it addresses is what Google calls an N by N integration bottleneck, where every business has to build a custom connection for every AI surface. UCP collapses that into one integration covering discovery, cart and checkout.

The mechanics are worth knowing even if you are years from implementing. Businesses publish a JSON manifest at /.well-known/ucp so agents can discover their capabilities and payment options without a hard-coded integration. The standard separates capabilities such as checkout and product discovery from extensions such as discounts, separates payment instruments from payment handlers, supports REST, MCP and A2A as transports with compatibility for AP2 agentic payments, and uses tokenised payments and verifiable credentials between agents and business systems.

Google's AI optimisation guidance already nods at this, recommending that site owners consider agent-friendly design for browser agents and naming UCP as an emerging protocol. Read in the right order, that is a signal about where the work goes next.

What it means in practice for a UAE business. If you sell online, your product, price and availability data is becoming an interface rather than a page, and the quality of that data decides whether an agent can transact with you at all. If you sell services, the equivalent is whether an agent can determine what you do, where you operate and how to start an enquiry without a human interpreting your homepage. That is a question about data and structure, which is why this sits with our ecommerce development and web development teams rather than with a content plan.

What to do in the next 90 days

No replatforming, no rewriting your site for robots. Seven things, in this order, all of which hold their value whatever happens next.

1. Check your generative AI inclusion setting in Search Console. Inclusion is the default, but confirm nobody changed it.
2. Open the Generative AI performance report and record a baseline. If there is no data, that is information too.
3. Fix the retrievable layer first. Indexing, rendering, canonical consistency, page experience. Nothing else works without it.
4. Audit your entity consistency. One name, one address, one phone number, one description, everywhere. This is a day of work with disproportionate effect.
5. List the twenty sub-questions your sales team answers every week and check which have a clearly headed answer on your site. Write the ones that do not.
6. Add original substance to your best commercial pages. Your own data, your own project detail, your own named experts. This is the layer competitors cannot copy.
7. Get your product, service and location data accurate and structured, because that is the groundwork for the agentic layer whether or not you ever implement a protocol.

Notice what is not on that list. No llms.txt deliverable, no content chunking project, no AI rewrite, no purchased mentions. Those are the five things Google says you can skip, and skipping them frees the budget for the work that matters. Our SEO services team runs this sequence with the development team in the same office, which matters here because four of those seven items are build work rather than content work.

Frequently asked questions